Unified Spatio-Temporal Encoding and Noise-Robust Dual Diffusion for Anomalous Edge Detection in Dynamic Graphs

Hang YuOpenReview 2026 · OpenReview

Abstract

Detecting anomalous edges within dynamic graphs is a cornerstone task for securing social networks, financial systems. Despite its importance, current state-of-the-art methods frequently struggle to capture comprehensive spatio-temporal dependencies and fail to effectively model anomalies due to noisy supervision signals. In this work, we present a novel framework designed to overcome these limitations through two key components. First, we introduce a spatio-temporal graph unified representation(STGUR) module. By leveraging unified spatio-temporal encoding alongside position-aware attention, this module synthesizes highly expressive edge representations that preserve complex structural evolutions. Second, to tackle the challenge of unreliable training data, we propose a calibrated dual diffusion anomaly mechanism. This framework utilizes a specialized calibration loss to rectify the inherent noise found in pseudo-anomalous labels, thereby enabling more discriminative anomalous edge modeling. Extensive empirical studies on multiple real-world datasets demonstrate that our approach consistently outperforms existing baselines.
BibTeX
@misc{yu2026unified,
  author = {Hang Yu},
  title = {{Unified Spatio-Temporal Encoding and Noise-Robust Dual Diffusion for Anomalous Edge Detection in Dynamic Graphs}},
  year = {2026},
  url = {https://openreview.net/forum?id=M7DvvYOE2c},
  howpublished = {OpenReview},
}

Loading related content...

Figure